proof

Reference workflow and artifacts
Not claims. Outputs.
Reproducible on independent machines.
Reference workflow

This workflow is intentionally concrete:

capture →
inference trace →
commitment →
packaging →
offline verify →
deterministic replay →

Capture
Sensor streams are captured and deterministically segmented into fixed windows (“Segments”).

Inference trace
When analytics are used, inference artifacts include model identity and execution context identity and are bound to the exact input Segments consumed. A verifier traces outputs back to segment digests and timestamps and flags context mismatches.

Commitment
For each Segment, a cryptographic digest is computed during capture and chained to the prior Segment. A signed Manifest records boundaries, digests, continuity links, time source metadata, and key identifiers.

Packaging
Segments, the signed Manifest, provenance objects, and verification material are assembled into a .loko package suitable for transfer and case attachment.

Offline verify
On an air-gapped machine, a verifier recomputes digests, validates chain continuity, validates Manifest signatures, detects missing content, and produces a pass/fail report with specific error reasons.

Deterministic replay
A replay client reconstructs review timelines deterministically from Manifest boundaries. Independent machines reproduce identical segment boundaries and discontinuity markers. Replay and export events are recorded in an integrity-protected audit and handling ledger.

.loko package

Portable structure that survives transfer and audit.

# .loko package example
<session_id>.loko/
  manifest/
    manifest.json            # signed boundaries, digests, continuity links, time source, key IDs
    manifest.sig
  segments/
    seg_00000.mkv            # fixed window media segment
    seg_00001.mkv
    ...
  provenance/
    capture_context.json     # device, config, time source declarations
    inference_context.json   # model + runtime identity when analytics are used
  verification/
    public_keys/             # verification material for signature validation
    verifier_report.json     # pass/fail + explicit reasons
  ledger/
    handling_ledger.json     # replay/export/policy outcomes, integrity protected
Verifier Report

This is what independently verifiable looks like.

{
  "package_id": "…",
  "verified_at": "…",
  "verifier_identity": "…",
  "result": "PASS",
  "checks": {
    "manifest_signature": "PASS",
    "segment_digests": "PASS",
    "chain_continuity": "PASS",
    "missing_content": "PASS",
    "time_source": "DECLARED",
    "key_policy": "OK"
  },
  "replay": {
    "boundary_hash": "…",
    "discontinuities": []
  },
  "errors": []
}
RELEASE GATE FLOW

Customer-run execution.
Loko validation.
Offline verification.

What you send
#Minimum receipt bundle

loko_receipts_<run_id>.zip

loko_run_summary_<run_id>.json

loko_manifest_<run_id>.json



#Optional (only if approved and needed)

loko_failures_<run_id>.json

loko_slices_<run_id>.json
Data minimization:
By default we request sanitized, non-production, non-PII artifacts. We do not request raw video/images, full labeled datasets, or model weights unless explicitly cleared.
Size targets:
Typically tens to hundreds of KB, occasionally a few MB. Default cap per run: 250 MB unless explicitly expanded.
Transfer options:
1. Time-expired upload link scoped to run ID (short expiry, constrained filenames/types, prefix isolation, logged access, retention window).

2. Encrypted bundle (customer-controlled encryption, key out-of-band, checksum included).
What Loko returns (post-validation)
  • Release Certificate (PASS/FAIL, gate ID, run ID)
  • Release Report (offline, self-contained HTML)
  • Improvement Recipe (in report or separate)
  • Updated package containing the above plus validated receipts and manifests
Independent re-verification
Any authorized reviewer can verify offline and reproduce the PASS/FAIL determination without relying on Loko or a live system.
Downgrade states

When systems cannot meet the highest bar, usability remains but the trust level must be explicit and governable.

EVIDENCE: passes offline verification and replay determinism

USABLE, NOT EVIDENCE: incomplete record or unverifiable components

ANALYTICS ONLY: inference present without committed integrity predicate
Procedural Legitimacy

On-record participation across standards, regulators and evidentiary rulemaking.

Receipts and acknowledgements are maintained.

download control plane pdf
U.S. Federal Courts
Evidence Rules submissions (FRE 707)
Written Testimony and Hearing Witness on Jan 15, 2026.
Filed
Dec 2025–Jan 2026

U.S SEC
Public comments (File 4-698)
Filed
Jan 2026

NIST & NCCoE
Public comments: 8587, 8596, NIST-2025-0035, AI Agent Identity & Authorization
Filed
Dec 2025–Feb 2026

California DMV - Autonomous Vehicle Regulations comments:
OAL File No. 2025-0415-04
Filed
Feb 2026

International & State-level Records
Notices submitted to multiple state court bodies (including FL, NY, CA, IL, TX) and EU AI Act.
Dec 2025–Jan 2026